Abstract
In this paper we propose a new method for land cover classification in hyperspectral remote sensing images by combining band selection with weighted spatial-spectral feature fusion. Spectral information for each pixel is represented by a spectral curve over all the bands. Spatial information is represented by a Bag of visual Words model within a small region around each pixel. A cluster-based band selection method is used before spatial feature extraction to reduce the computation complexity. Then spectral and spatial feature weights are learnt under a Support Vector Machine framework, obtaining a balance between the two basis features for each class. Classification results on three popular hyperspectral remote sensing images demonstrate that the proposed method can yield a higher accuracy and a lower false alarm rate compared with the other similar classifiers.